Models#Benchmark#Model inference
Mistral Large 4: Europe goes trillion
A 1T-parameter (49B active) MoE trained on 3,800 Blackwells in Europe; open weights this month and a €3B Series D.

Mistral shipped Mistral Large 4 in public preview on October 6, unofficially “ML4” and very officially “Le Chonk”: a natively multimodal hybrid instruct-and-reasoning MoE with 1 trillion total parameters and 49 billion active, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters, with the RL run generating about 33 billion tokens per day. It is Europe’s first front-rank answer to the US and Chinese flagships, and the company’s positioning is unsubtle: “#1 in Sovereign AI.” The training footprint deserves attention on its own — 3,800 top-tier accelerators is a serious infrastructure commitment, evidence that the efficiency route ends in the same place as everyone else’s: enormous capital expenditure, dressed in better unit economics.
Key points
- Architecture: 1T total / 49B active MoE, hybrid instruct-and-reasoning, multimodal input
- Training: 3,800 Grace Blackwell GPUs, entirely in Mistral-owned European datacenters
- Coding: 49.8% on the combined Coding Agent Index, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max
- Blind eval: second of five on Surge AI’s human eval (3.74), behind Claude Opus 5 (4.22)
- Security: 82% on CyberGym-E2E (“the highest of any model”), resists 93.3% of Lakera B3 attacks
- Openness: weights by month’s end (red-team first); pricing $1.36 in / $4.18 out per million tokens
- Capital: a €3 billion Series D — the largest European tech equity round on record
What the benchmarks say
Second place in blind eval says the top tier (Claude Opus 5) still has daylight; leading DeepSeek and Qwen on coding and agent benchmarks says the open-weight order is reshuffling beneath it. The caveat printed in every Mistral chart applies to its own numbers too — these are launch-day, vendor-run scores, and theurge to treat them as settled should wait for the weights. The Decoder reads the deeper move correctly: Mistral is selling the work the US closed models decline — top-of-table cyber benchmarks, open-source-best legal agent scores — positioning as the European option for security-sensitive industries, in the same “safety as a market” season as Anthropic’s catastrophe-risk prospectus language and OpenAI’s textGrain compliance watermark.
Sovereignty, with a balance sheet
The €3 billion Series D funds a route that is expensive but buys two things US labs cannot offer: data sovereignty for European enterprises with compliance demands, and supply-chain independence from US export policy. Set beside DeepSeek’s ¥8 billion round and the American trillion-dollar narratives, the frontier race now runs on three distinct capital-regulation operating systems.
The European route, priced in
The EUR 3 billion Series D is not “another OpenAI” money — it buys Europe a self-sufficient sovereign-AI supply chain: training on domestic soil, weights that can be self-hosted, pricing in euros. Governments and defense contracts in the UK, France and Germany are the first customers for that story, and the strength on HarveyAI’s legal benchmark and the cyber suites aims squarely at finance, legal and critical-infrastructure buyers who cannot ship data across borders. It is a narrower go-to-market than the American labs’ consumer land-grab — and a defensible one.
The real test: the weights
Preview benchmarks are vendor-reported; third-party reproduction only becomes possible when the weights land — and trust in launch-day numbers is depreciating fast, between Jev’s anti-benchmark stance and every leaderboard scandal since. For European enterprises, Mistral finally offers a frontier option that does not require leaving the jurisdiction; for the open-weight camp, the 1T/49B ratio signals that the MoE arms race has entered its stack-total, spend-active phase — 97% of the parameters idle until called, which is efficiency and a bet at once: the trillion dormant parameters pay off only under long-context and multimodal load. When the weights open, the $1.36/$4.18 pricing against DeepSeek and Qwen will decide whether the European flagship reads as a usable option or a sovereignty symbol.